Plant functional type mapping for earth system models

. The sensitivity of global carbon and water cycling to climate variability is coupled directly to land cover and the distribution of vegetation. To investigate biogeochemistry-climate interactions, earth system models require a represen-tation of vegetation distributions that are either prescribed from remote sensing data or simulated via biogeography models. However, the abstraction of earth system state variables in models means that data products derived from remote sensing need to be post-processed for model-data assimilation. Dynamic global vegetation models (DGVM) rely on the concept of plant functional types (PFT) to group shared traits of thousands of plant species into usually only 10–20 classes. Available databases of observed PFT distributions must be relevant to existing satellite sensors and their derived products, and to the present day distribution of managed lands. Here, we develop four PFT datasets based on land-cover information from three satellite sensors (EOS-MODIS 1 km and 0.5 km, SPOT4-VEGETATION 1 km, and ENVISAT-MERIS 0.3 km spatial resolution) that are merged with spatially-consistent K ¨ oppen-Geiger climate zones. Using a beta (ß) diversity metric to assess reclassiﬁcation similarity, we ﬁnd that the greatest uncertainty in PFT classiﬁcations occur most frequently between cropland and grassland categories, and in dryland systems between shrubland, grassland and forest categories because of differences in the mini-mum threshold required for forest cover. The biogeography-biogeochemistry DGVM, LPJmL, is used in diagnostic mode with the four PFT datasets prescribed to quantify the effect of land-cover uncertainty on climatic sensitivity of gross primary productivity (GPP) and transpiration ﬂuxes. Our results show that land-cover uncertainty has large effects in arid regions, contributing up to 30 % (20 %) uncertainty in the sensitivity of GPP (transpiration) to precipitation. The availability of PFT datasets that are consistent with current satellite products and adapted for earth system models is an important component for reducing the uncertainty of terrestrial biogeochemistry to climate variability.


Introduction
Terrestrial biogeography holds a key role in determining spatial patterns of biogeochemical cycling (Olson et al., 1983), biodiversity (Kleidon et al., 2009) and the consumption and production of natural resources (Foley et al., 2005;Haberl et al., 2007). An improved understanding of global biogeography is required to provide a baseline for assessing the vulnerability of the carbon and water cycle and other ecosystem processes related to ongoing global change. Such baselines are also relevant to earth system modeling research, especially for dynamic global vegetation modeling (DGVM), with applications extending to model initialization, optimization, and benchmarking (Plummer, 2000). However, available remotely-sensed datasets for land cover show large variability (Giri et al., 2005), partly due to differences in data retrieval (i.e., satellite properties) and partly because there is no standard approach to classifying continuous vegetation cover into discrete categories. In addition, to be comparable to plant functional type (PFT) definitions used by DGVM models, land-cover legends must be cross-walked (reclassified) to broader PFT categories (Jung et al., 2006). To address the variability between land-cover products and the challenges of cross-walking land-cover legends, we developed a methodology to process an ensemble of PFT datasets corresponding to the most common global land-cover products available. In our analysis, the pattern and drivers of variability across this ensemble is quantitatively assessed and attributed to interpret the effects of land-cover uncertainty on biogeochemical fluxes.
Published by Copernicus Publications on behalf of the European Geosciences Union. There are now several (Table 1) moderate resolution global land-cover datasets available from different satellite sensors and research groups (Friedl et al., 2002(Friedl et al., , 2010Bartholome and Belward, 2005;Arino et al., 2008) providing an opportunity to assess ensemble variability. Although these landcover datasets provide new opportunities for model-data assimilation studies to assess the effects of land-cover feedbacks (Quaife et al., 2008;Sterling and Ducharne, 2008;Jung et al., 2007), their approach for classifying land cover is not yet consistent with Earth System Model (ESM) requirements. This is because the concept of plant functional types used in ESMs cannot be mapped directly using remote sensing data since PFT traits represent a combination of spectral relationships, and climatic, ecological, and theoretical assumptions (Smith et al., 1997;Sun et al., 2008;Running et al., 1995;Ustin and Gamon, 2010). The PFT concept consists of aggregating multiple species traits, allowing for the reduction of thousands of species to a small set of functional groups (typically <15) defined by their phenology type, physiognomy, photosynthetic pathway, and climate zone. The advantage of the PFT classification system is that it allows the possibility for posing testable hypotheses that are feasible at global and centennial scales (Smith et al., 1997).
Existing PFT datasets include those by Bonan et al. (2002) for the Community Land Model, with updates from (Lawrence and Chase, 2007), by Verant et al. (2004) for the Orchidee DGVM, and by Lapola et al. (2008) for the SSiB model. Improvements to these PFT datasets are currently needed to expand the availability of land-cover datasets to allow consistency with a more complete set of satellite sensors and more detailed or revised climate zone data, and to take into account current human land-use patterns. For example, Bonan et al. (2002) used multiple data sources to combine the IGBP-DISCover Global Land Cover Classification data (IGBP GLCC) and phenology-type data (from 1992-1993 AVHRR data) with vegetation continuous fields from De-Fries et al. (2000). They assigned biome types from bioclimatic definitions provided by Prentice et al. (1992) based on gridded climate data from Legates and Wilmott (1990), creating one of the first ESM-relevant PFT legends (Table 2) for the Community Land Model 3.0 (Dickinson et al., 2006). In comparison, Verant et al. (2004) combined simplified Olson biomes with IGBP GLCC data to create a PFT map for the Orchidee DGVM (Krinner et al., 2005). Lapola et al. (2008) developed a global PFT map by reclassifying legends from Olson et al. (1983) and Matthews (1983) and filling areas of mismatch with regional land cover information. A different PFT legend accompanies the MODIS land cover product using categories defined by Running et al. (1995) and has been developed from GLC2000 (Wang et al., 2006). For these particular PFT legends, the classifications include phenology type but not the associated climate zone, which is needed to assign climate-specific physiological parameters to each PFT (i.e., Sitch et al., 2003). As a consequence, vegetation models using these particular PFT datasets must assume that biochemical and biophysical PFT parameters are constant globally across different climate zones (e.g., see Alton et al., 2009).
Our study contributes to ESM and DGVM modeling by developing four global PFT datasets (Table 1) using a consistent methodology applied to satellite data that vary by sensor (spatial and spectral resolution), classification system, and time period. The Köppen-Geiger climate classification scheme is used to associate physiognomy and phenology type with climate zone, and the pattern of uncertainty among the four classification systems is evaluated using a beta diversity metric. We provide an example of the importance of land-cover uncertainty on land-surface climate sensitivity by prescribing vegetation types and analyzing biogeochemistry with the LPJmL DGVM. In this experiment, we quantify the sensitivity of water and carbon fluxes to climatebiogeochemical fluxes highly modified by human activities (Oki and Kanae, 2006;Schimel et al., 2001) -and evaluate how land-cover uncertainty alters these relationships.

Land cover and climate zone datasets
Land-cover datasets, described in Table 1, were manually reclassified to PFT specific phenology type and physiognomic categories. The resulting categories were merged with climate zones defined by the Köppen-Geiger classification system to resolve to PFT classes. The merged dataset was aggregated to 0.5 • spatial resolution (corresponding to the climate and soils data used in LPJmL), representing the fractional abundance of PFT mixtures within a grid cell. All analyses were conducted at the global scale in Plate-Carrée (WGS84) projection, area correcting grid cells during post-processing when necessary. The original land-cover datasets varied in spatial resolution, time period of data collection, classification approach, and accuracy and are discussed below. The Köppen-Geiger dataset was created by Peel et al. (2007) from over 4000 metrological stations contained in the Global Historical Climatological Network v2.0 database. The authors calculated climate indices (i.e., seasonal means, minimums, and maximums) for the stations from precipitation and temperature for their entire time series (mostly, the 20th century) and then interpolated to a 0.1 • resolution grid (not accounting for elevation). These indices were classified into one of 32 possible climate zones (Table 3) according to the original Köppen-Geiger classification system (Köppen, 1936).
The GLC2000 land-cover data were generated from SPOT-VEGETATION (SPOT 4) and ATSR-2/DMSP sensors and are available for most of the vegetated surface of the globe (75 • N to 56 • S, excluding Antarctica) at 1 km resolu-tion (Bartholome and Belward, 2005;Hugh et al., 2004). The data were collected between November 1999 and December 2000. The GLC2000 classification (Table 4) was conducted by regional expert groups following an unsupervised classification of 19 similar geographic regions using the LCCS nomenclature (22 categories for global purposes).
The GlobCover data became available in 2008 (Arino et al., 2008) and represent the highest-spatial resolution data available for global extent at this time (0.3 km resolution). The classification system also follows the LCCS system (22 categories, Table 5) and the spectral data were acquired from the MERIS sensor on-board the ENVISAT satellite between June 2004 and December 2006. Individual pixels are classified using unsupervised and supervised approaches on sub-global regional clusters.
Two versions of the EOS-MODIS land cover data (MOD12Q1), V004 and V005, were used in the analysis. These differ in several aspects, including temporal coverage, spatial resolution, and classification methodology, but both use the same 17 IGBP categories (Table 6) (Friedl et al., 2010). These land-cover classes were categorized using a globally consistent supervised classification approach. V004 is available globally at 1 km resolution from data acquired in 2001 while V005 is available at 0.5 km resolution at annual resolution (starting in 2001). Both products have multiple legends available, and here we worked with the IGBP legend (Table 6), the primary MODIS legend from which the other legends are derived and most relevant for reclassifying to phenology categories (next section).  Peel et al., 2007) and their simplified equivalents required for DGVM PFT classification. Where T hot and T cold are temperature of the hottest and coldest month ( • C), and MAT is mean annual temperature ( • C).

Number Code
Köppen-Geiger biome category PFT biome equivalent (this study)

Reclassifying the legends
The land-cover data were first cross-walked (reclassified) to a phenology-based legend consistent with the plant functional types used in major DGVM and land surface models (Table 2). Figure 1 illustrates the flow of data processing, with the merging of the phenology type and climate data described in the following section. Manually reclassifying legends is inherently subjective, especially with the treatment of mixed vegetation categories where multiple possible classes must be reduced to only one class (Jung et al., 2006;Giri et al., 2005). The possible bias resulting from reclassification was handled, in part, by comparing levels of agreement among the aggregated PFT classification datasets. The comparison highlighted the spatial pattern of dissimilarity and provided a means for qualitatively evaluating the contribution of uncertainty from reclassification problems and from actual remote sensing differences. The reclassification was conducted on the original resolution of the land-cover dataset and implemented in C programming language. Artificial surfaces and associated areas Urban 23 No data No data Fig. 1. Illustration of the flow of analysis and the reclassification of the climate zone data to simplified biomes, and the land-cover data to their phenology types, physiognomy and land-use counterparts. Tables 4, 5, and 6 list the original land-cover classes from GLC2000, GlobCover, and MODIS and their corresponding reclassification into phenology type (Table 2). Six PFTspecific phenology-type/physiognomy classes were predefined, corresponding with categories used in several DGVM models (broadleaf evergreen BrEv, broadleaf deciduous BrDe, needleaf evergreen NeEv, needleleaf decid-uous NeDe, natural grassland NatGrass, and managed grassland ManGrass (representing either pasture or crop)). For some categories, the reclassification was straightforward, i.e., GLC2000 "Tree cover, broadleaf, deciduous, closed" was reclassified to "broadleaf deciduous." Other categories, for example, GLC2000 "Tree cover, mixed leaf type," or MODIS IGBP "Open shrublands" did not correspond to a Geosci. Model Dev., 4, 993-1010, 2011 www.geosci-model-dev.net/4/993/2011/ single PFT phenology/physiognomy class. In these cases, the land cover class was reclassified to one of several possible phenology-types and physiognomy classes whose probability was assigned by assessing the supplementary data regarding the legend definitions or examining the spatial pattern of observed land cover classes, and based on expert opinion on how the class might be composed of various phenology types (similar to Wang et al., 2006). In these cases, for example, a "mixed tree cover" category would yield 25 % equal probability (using a uniform distribution for all mixed land cover categories) with the grid cell being reclassified to either BrEv, BrDe, NeEv, or NeDe. This approach resulted in a single category cell, but when the cells were aggregated to coarser resolution (described below), the relative PFT fractions more realistically represented the original mixed forest classes (for example, aggregating from 1 km mixed forest category to 0.5 degree resolution results in 0.5 degree fractions equal to 0.25 for BrEv, BrDe, NeEv, and NeDe, summing to 1.0 for an aggregated cell).

Merging and aggregating phenology and climate zones
The Köppen-Geiger dataset was first adjusted to expand its coastal grid cell definitions to neighboring ocean grid cells to allow a complete overlay of land cover with climate zone. The buffered Köppen-Geiger data were then downscaled to the spatial resolution of the corresponding land-cover dataset using a nearest neighbor resampling algorithm. The resampled Köppen-Geiger data were reclassified into one of three major biome types (following the rules described in Table 3), namely: tropical, temperate and boreal. The temperate and the boreal biome were further subdivided into either cool (<22 • C) or warm (=>22 • C) types to distinguish between C3 or C4 photosynthesis in the former, and temperate needleleaf and broadleaf trees in the latter (based on their PFT temperature establishment thresholds in Table 2). While C4 grasses can establish at cooler temperatures (i.e., the LPJ model uses a temperature of 15 • C, Table 2), this temperature threshold (22 • C) has been shown in prior studies to be a critical "crossover" temperature for C3 and C4 adaptations (Collatz et al., 1998). Each of the 4 reclassified phenology type datasets were then merged with the climate zones to produce the final PFT classification at the spatial resolution of the original land cover data following the assembly rules in Table 7. Some exceptions were made to account for the full combination of phenology and climate zone possibilities. For example, because there are few to no deciduous needleleaf PFTs observed in tropical and temperate ecosystems, this phenology type was treated as tropical broadleaf raingreen (deciduous) or temperate broadleaf summergreen PFT. Natural and managed grasslands were split into the C3 and C4 photosynthetic pathways according to temperature thresholds that defined tropical versus temperate, and cool versus warm temperate biomes from the Köppen-Geiger data. This approach may underestimate C3/C4 grass mixtures or C4 summer crops (i.e., maize) that might be planted in cooler regions (Ramankutty and Foley, 1998).
The PFT classifications were aggregated to a spatial resolution of 0.5 • by summing the area of each PFT class within the corresponding 0.5 • cell (16 classes, Table 7) and dividing by the grid cell area. A spatial resolution of 0.5 • was chosen for this study because most models in the ESM community use climate and other ancillary driver (e.g., soil type) data at this resolution, or greater (Zobler, 1986;New et al., 2002). The aggregation of PFT fractions can also be carried out at finer resolution, but at smaller window sizes the estimates of fractional PFT coverage may become more sensitive to the selection of probability distribution. Each of the four PFT fractional abundance files were filtered with a global land/water mask, which was derived from a global soils database (Zobler, 1986). This ensured that the terrestrial surface area and land/ocean boundaries were equal between datasets.

Measuring PFT agreement
We analyzed the agreement between PFT fractional abundance (and re-groupings of PFTs by various traits) with a beta (ß) diversity metric (mean Euclidean distance) calculated for each grid cell. Euclidean distance is a measure of dissimilarity between groups with multiple members (Legendre et al., 2005) and is commonly used to summarize landscape species diversity from multiple sampling plots (Whittaker, 1972). In our case, the "plots" were the grid cells which contained the fractional PFT abundances contained from the different classification datasets. This analysis had two objectives; the first was to assess, geographically, where regions of high uncertainty in PFT abundance may exist, the second was to help evaluate the methods for the reclassification of legends, especially for the mixed vegetation categories.
The beta diversity metric was calculated for each grid cell for each of the four datasets, for the standard PFT classification, and for three re-groupings based on PFT traits. These regroupings were 1. Phenology type (total evergreen versus total deciduous fraction), 2. Physiognomy (total woody versus total herbaceous fraction), and 3. Management status (natural grass versus managed grass). Equation (1) presents the variables used for calculating the Euclidean distance, the mean of which, we consider to represent beta diversity, ß. For every grid cell c, the Euclidean distance, D was calculated between every combination of classifications, N (1. . . 4) composed of 10 PFTs (I = 10) and their corresponding fractional abundance Afor the different classifications (j and k).
www.geosci-model-dev.net/4/993/2011/ Geosci. Model Dev., 4, 993-1010, 2011 Table 6. Modis collection 4 and 5 legend (IGBP) and corresponding DGVM phenology class (from Table 2). For mixed phenology cells, it was assumed that any phenology type could be found, these were determined randomly using a uniform distribution.  The mean Euclidean distance (beta diversity) between groups (D c ) was calculated as the mean of the diagonals from the resulting matrix to represent overall dissimilarity (Legendre et al., 2005;Whittaker, 1972). The variance of the Euclidean distance matrix was also calculated using the same approach, but taking the variance of the matrix diagonals rather than the mean. The two approaches were used to aggregate the distance matrix into a single index as recommended by Legendre et al. (2005). ß was plotted as both geographic maps and as latitudinal summaries to explore the spatial patterns of uncertainty for each grouping. An ordination analysis was conducted for four individual grid cells representative of major biomes (temperate, tropical, boreal, and desert) to investigate the similarity between PFT products and to display the main gradients partitioning them. The PFT products included those developed in this study, the existing products described in the Introduction for Orchidee and CLM, and the results from a DGVM simulation from LPJmL (described below). Non-metric multidimensional scaling (NMDS) was chosen (using a Euclidean distance matrix); this ordination method is less sensitive to non-linear relationships among variables.

Prescribing PFT data to a dynamic global vegetation model
PFT fractions were prescribed to the LPJmL DGVM, an ecosystem model that simulates global biogeography and biogeochemistry via coupled water-carbon cycling and vegetation dynamics (Sitch et al., 2003  were removed to allow the establishment of PFTs wherever they were prescribed from the external datasets. In the LPJmL model, diagnostic PFT fractions replaced the variable for maximum annual fraction of photosynthetic absorbed radiation (FPAR) while not modifying the vegetation dynamics or physiology modules. Monthly climate data (precipitation, temperature, cloud cover, and wet days) and annual CO 2 concentrations were prescribed from the Climatic Research Unit TS3.0 dataset (Mitchell and Jones, 2005) and the Carbon Dioxide Information Analysis Center (CDIAC). Following a 1000-yr spin up to equilibrate vegetation and carbon pools, a transient simulation, including the effects of prognostic fire  Table 2. (Thonicke et al., 2001), was initiated beginning in 1901 and ending in 2005. Managed grasslands were treated as in Bondeau et al. (2007), with harvest occurring repeatedly during the year when peak leaf area index (LAI) was reached. Annual GPP and transpiration (from 1956-2005) were regressed with mean annual temperature, total annual photosynthetic active radiation (PAR), and total annual precipitation, from the same time period, to calculate partial correlation coefficients used to interpret the sensitivity of the biogeochemical fluxes to climate.

Uncertainty in global PFT fractions
At the global scale, managed and unmanaged grassland PFTs, which includes croplands and pasture, were most abundant in terms of percent land area (∼30 %), followed by tropical (∼15%), boreal (∼12 %), and temperate PFTs (∼11 %) ( and tropical evergreen (TrEv) PFT fractional coverage also had high uncertainty, both in savanna regions, and in the Amazon and Congo River Basins (acronyms explained in Table 2). The lowest uncertainty was found for the boreal needleleaf PFT (BoNe) and natural/managed C3 grasslands.
The spatial uncertainty of PFT agreement was low in the boreal biome, increasing to some extent in the temperate biome, and highest in the tropics (Fig. 2). The disagreement between land-cover datasets was greatest for deciduous PFT phenology types compared to evergreen PFTs, which was especially important in the tropics where the GlobCover dataset categorized the Amazon and Congo River Basins as mixedleaf or seasonal forest (in comparison to the "evergreen" phenology type for the other datasets). Managed and natural C3 grasslands showed high dissimilarity in the mid-western United States mostly because GlobCover had low fractions of managed grassland in this region ( Fig. 2; MGrassC3 low abundance between 40-50 • N). C4 grasslands (both natural and managed) showed high dissimilarity because of differences related to distinguishing natural versus managed grassland, and because of differences in how the land-cover datasets treated barren/dry areas. For example, much of interior Australia was classified as "open shrublands" by the IGBP MODIS legend ( Fig. 3a; higher abundance of TrEv and TrRg compared to Fig. 3b). Our reclassification for the IGBP legend assigned 40 % of this class to woody PFTs, 20 % to grass PFTs and 40 % to barren. In comparison, the LCCS legend reclassifies part of this region as "sparse vegetation", which we reclassified as 40 % grass PFT and 60 % bare, consistent with the supporting documentation for LCCS ( Fig. 3b; lower abundance of TrEv and TrRg). The IGBP "open shrubland" category also includes tundra and permafrost biomes (in addition to the warmer arid regions). Herbaceous cover may be higher in these cooler regions than in "open shrublands" of warm regions ( Fig. 3b; NatGrassC3 replaces by BoNe), which suggests that further refinement of shrubland categories could improve differences between land-cover products.
Spatial resolution and the detail of land-cover categories had important effects in intensively managed landscapes. For example, in the Southeastern United States, the Glob-Cover dataset (0.3 km) and LCCS legend (22 classes) bet-ter distinguished secondary succession vegetation ( Fig. 3b; higher abundance of TeNe) (i.e., pine forests from agricultural abandonment, Christensen and Peet, 1984). Climatic differences across small gradients were generally detected by the Köppen-Geiger classification, despite elevation not being included in the interpolation process. These topographic features were apparent in the north-south divide along the island of Madagascar (Fig. 4a-d; NatGrassC4 versus TrRg), along the Andes, separating the Amazon rainforests from highelevation grasslands, as well as in Ethiopia, where the effects of the highland rift-valley corresponded to C3 grasslands in a region mostly surrounded by C4 climate zones (Fig. 4a-d;  NatGrassC3). GlobCover and GLC2000 detected higher fractions, ∼12-13 %, of managed grasslands globally. Compared to the MODIS products, this estimate is slightly less than ∼14 % for cropland area found in previous studies scaled to FAO country statistics (Ramankutty and Foley, 1998;Klein Goldewijk et al., 2007). GlobCover also distinguished alpine vegetation communities in more detail (compared to GLC2000 and MODIS) for the European Alps. C4 abundance was higher than in Still et al. (2003) who found that globally, C4 vegetation compose ∼15 % of terrestrial vegetation (whereas the estimates presented here are closer to ∼22 %, Table 8). The higher estimates for C4 grass abundance are due, in part, to differences in the IGBP "grasslands" and LCCS "sparsely vegetated" categories, which corresponded to the LCCS "sparse vegetation" or "barren" categories ( Fig. 3a and b; see MGrassC4). For IGBP, "grasslands" were reclassified to 100 % grass cover (Table 6), but for LCCS "sparse vegetation," only 40 % grass cover (Table 5).
The disagreement between PFT trait-groups (see categories described in Methods: Measuring PFT agreement), was highest in the tropics (Fig. 5; 20 • N to 20 • S) and for a "hotspot" in the mid-western United States ( Fig. 5; 35-55 • N) resulting from managed versus natural grassland classification. Phenology-type disagreement was also high in the tropics, but in general, structural (physiognomy) observations appear to have high agreement (Fig. 5). In the northern temperate zone, grasslands were more consistently classified as managed, with the exception of mid-western USA, where GlobCover underestimated cropland fraction compared to GLC2000 and MODIS ( Fig. 2; MGrassC3). Tropical savannas and warm-climate croplands emerged as bands of disagreement, because of differences between the IGBP and LCCS classification for natural and managed C4 grassland and shrubland categories. A notable region of high uncertainty was the Karakum desert in Central Asia which was classified as "barren" or "sparse vegetation" in LCCS and as "grassland" in IGBP leading to large differences in estimated PFT fractions (Fig. 6).

Land cover and the uncertainty of fluxes to climate
In diagnostic mode (with PFT distributions prescribed), global GPP ranged from 130.9 to 134.9 PgC a −1 (averaged over 1996-2005) and transpiration ranged from 43 200 to 44 600 H 2 O km 3 a −1 . These global values were similar to the prognostic (dynamic vegetation) simulation, using the Hyde dataset for managed grasslands (Klein Goldewijk and Batjes, 1997), which produced values of 131.0 PgC a −1 and 39 000 H 2 O km 3 a −1 . All estimates are close to previous analyses of global carbon Beer et al., 2010) and water fluxes (Gerten et al., 2005). GPP and transpiration sensitivity to climate followed similar patterns observed in previous studies (Nemani et al., 2003), with temperature important in northern latitudes, radiation more limiting in the wet tropics, and precipitation a dominant feature globally (Fig. 8a). As in Beer et al. 2010), precipitation was the most important global climate variable controlling GPP (65-70%) and transpiration (58-63 %). The range of uncertainty was similar for either GPP or transpiration sensitivity to climate; with agricultural regions in mid-western USA and Europe, and arid regions in Australia and S. Africa showing high uncertainty in the sensitivity of GPP to precipitation (Fig. 8b). In agricultural regions, the lower fractional coverage of croplands in the GlobCover product led to higher grassland LAI (because of no harvesting), causing higher sensitivity (or correlation coefficient) to precipitation. In semi-arid regions, the MODIS products led to higher GPP sensitivity to precipitation (Figure 8) because of a higher abundance of woody species (with deeper rooting strategies) unable to compete efficiently for minimal rainfall with grasses that had shallow rooting strategies.

PFT datasets and themes for improvement
PFT datasets must remain consistent with available satellite products used for data assimilation (i.e., LAI, FPAR or biomass assimilation) and account for recent changes associated with land-use dynamics. Here, we have evaluated an approach for establishing a series of PFT datasets for use within Earth System Models and DGVMs, and explored patterns of disagreement and their propagated effects to land-surface process uncertainty. To date, available PFT datasets for ESM models have been limited to single satellite sensors and partly outdated land-cover data (i.e., 1992(i.e., /1993. The approach used in this study increases the resources available for evaluating ensemble uncertainties introduced from landsurface state variables and we discuss possible opportunities for refining classification methodologies. Fig. 6. TMap of mean fractional dissimilarity (0 is complete agreement and 1 is complete disagreement) considering all PFTs using the beta-diversity metric (Eq. 1). Representative grid cells shown in Fig. 7 illustrate the main patterns of clustering and spread of the existing products in comparison to the new PFT products presented here. Fig. 7. Comparison of the 4 new PFT products to one another and the existing products based on older datasets. The NMDS ordination shows the degree of differences and clustering of the products (black names) and the major PFT gradients (p < 0.05) that explain the differences in the gradients (red names). The LPJ product is the PFT composition from a fully dynamic vegetation simulation, described in the Methods. The main areas of disagreement were found in regions of intensive land management or in regions where accurate spectral discrimination between similar land cover, but different land use, caused classification problems (e.g., cropland and natural grassland). High-disagreement was also found in warm/cool arid regions where land-cover categories were either too broadly defined or because different legend types had conflicting tree-cover thresholds to define forest vegetation. Previous studies have also observed that dryland systems, which include heterogeneous mixtures of grasslands, croplands and savanna shrublands, feature as prominent zones of disagreement (Giri et al., 2005;Herold et al., 2008;Fritz and See, 2008). Our results confirm that this disagreement scales to PFT groupings and contributes a large part of the land-surface process uncertainty for GPP and transpiration sensitivity to climate (Fig. 9).
Much of this disagreement results from differences in the classification for forest land; the LCCS definition for forest is an area with more than 15 % tree cover, whereas IGBP uses a 60 % tree-cover threshold. Consequently, the MODIS product has a much larger fraction of (non-forest) shrubland and savanna systems, which are categorized as various "open" or "closed" forest types in GLC2000 and GLOBCOVER, Tables 4 and 6. Such problems stem from defining forest structure from forest cover, which may be overcome with new developments in satellite-based lidar, which can successfully provide measurements of tree height and vertical structure at global extents (Lefsky, 2010). Overlap of cool or warm aridland categories (i.e., grassland, shrubland, barren), also introduced error in deserts and tundra regions, where broadly defined categories could be improved by including climate information. Arid-regions have large global coverage and recent work suggests that these ecosystems have a significant influence on global biogeochemistry and the climate system itself via biophysical properties (Rotenberg and Yakir, 2010), hence a better understanding of their distribution could contribute to reducing uncertainty of global climate processes.
The humid tropics remain an additional challenge for biogeochemical modeling, in part, due to a need to better understand basic ecophysiological processes (Baker et al., 2008), but also because of data collection limitations. Remote sensing data for tropical vegetation must first be processed to remove seasonal cloud cover and aerosol contamination Kobayashi and Dye, 2005), and also account for the rapid pace and complex spatial patterns arising from deforestation and fire (Morton et al., 2006). In our study, these measurement problems contributed to land-cover uncertainty in the classification of natural versus managed grasslands in southeastern Brazil, and in the classification of the phenology type for the wet-tropical forests. Quantifying the degree of seasonal tropical forest phenology is a longstanding problem and studies disagree at the level to which these forests shed their leaves during the dry season (Samanta et al., 2010;Saleska et al., 2007). Despite the disagreement among tropical PFTs, the uncertainty in GPP and transpiration sensitivity to climate was relatively low (Figs. 8b and Fig. 9, Tropical America and S. Africa regions). This was because of similarities in the ecological functioning of the modeled raingreen and evergreen PFTs in non-water stressed environments (see ). Fig. 9. Cumulative partial correlation between GPP and precipitation using the same approach as described in Beer et al. (2010). Zones correspond to TRANSCOM 3 biome regions (Gurney et al., 2003), and legend colors red for GLC2000, green for GlobCover, light blue for Modis v4, and dark blue for Modis v5.
Our analysis suggests that the PFT uncertainties could be reduced by using land-cover data based on high to moderatespatial resolution and a larger number of legend categories (as in GlobCover). For example, the more detailed LCCS legend was better able to handle dry-land classifications than the coarser IGBP legend and GlobCover appears to classify heterogeneous landscapes well. Future versions of MODIS land-cover data are expected to include the LCCS legend (Friedl et al., 2010), which in addition, should further reduce errors from user-based reclassification necessary for landcover product comparisons.

Application with Earth System Models
Developments in ESM models will likely focus on including a higher diversity of PFT and crop functional types and entirely new approaches that include adaptive, rather than fixed, traits for PFT parameters (Scheiter and Higgins, 2009). The number of PFT types is not restricted to those presented in our study and the next generation of DGVMs, with higher-functional diversity, will be better prepared to evaluate ecosystem resilience and stability hypotheses related to global change. However, increasing PFT diversity requires detecting life history traits (i.e., growth form and dispersal rates) not readily observable at global scales or from space (Ustin and Gamon, 2010). The utility of the PFT approach for hypothesis testing and linkage to remote sensing will remain important. Finer resolution categories of crop types has been shown to be important for global biogeochemical cycling  but crop types or crop cover (or pasture) is not easily distinguished in the global land-cover datasets, highlighting the importance of integrated land-cover mapping approaches. Managed grassland categories can be subdivided using regional statistics on crop use, similar to methods described by Ramankutty et al. (1998), but many earth system models are at the early stages of incorporating crop functional types.
By forcing LPJmL with diagnostic PFT fractions we were able to illustrate the utility of ensembles of land-cover approaches and the application of diagnostic datasets. Interestingly, global estimates of GPP and ET were similar, regardless of land cover, confirming studies conducted at continental scales (Jung et al., 2007). However, we show that there are large regional differences, 20-30 %, in the sensitivity of biogeochemical fluxes to climate that are directly linked to land-cover uncertainty. High-PFT uncertainty did not always correspond to high biogeochemical cycling uncertainty (e.g., wet tropics), illustrating that propagated errors may differ from the initial condition agreement and that the choice of evaluation metric is important. These PFT datasets have applications beyond ESM modeling and can be integrated with bottom-up studies, include accounting methods for evaluating carbon stocks (Kindermann et al., 2008), or as base-maps that can inform biodiversity-patterns related to biogeography (Loucks et al., 2008).